Asked about Getting the most from the hardware
What does kernel fusion do?
An answer that works
Combines several operations into one kernel so intermediate results stay on-chip instead of being written to and read from memory between steps. The saving is in memory traffic rather than arithmetic.
Reading an answer is not being able to give one. The section that teaches this is Getting the most from the hardware, 4 hours of GPUs & accelerators.
Learn this properlyAlso asked about GPUs & accelerators
Why are GPUs used for deep learning?What is mixed precision training?Your GPU shows 30 percent utilisation. What is likely wrong?Is LLM inference compute-bound or memory-bound?How do you tell whether a GPU workload is limited by compute or memory bandwidth?You need to serve a 7B model cheaply at moderate traffic. How do you choose hardware?